A KDC error of KDC_ERR_SVC_UNAVAILABLE indicates that the client should 
try other KDCs.  Right now we do try other KDCs, but only after timing 
out on the KDC we received the reply from.

We handle KDC_ERR_SVC_UNAVAILABLE using the msg_handler callback 
argument of k5_sendto.  When service_fds receives a reply from the 
function, it invokes the callback, which decodes the reply as a KRB-
ERROR and returns 1 if it is a KDC_ERR_SVC_UNAVAILABLE error.  
service_fds reacts by continuing the loop, but does not mark the server 
as dead, so the loop continues to wait for a reply from the KDC (even 
though it has already replied) until the next timeout interval.

We can solve 90% of this problem by marking the server is dead when the 
msg_handler callback returns 1, so that selstate->nfds decrements to 0 
and service_fds returns to k5_sendto.  If we are still waiting for other 
servers, we will still continue waiting for them longer than we really 
need to, but that's also true if we got a communications failure talking 
to the most recently resolved server.